9 Ways to Extract Text From an Image (OCR Methods Compared)

Optical character recognition turns pixels back into text. The technology is mature enough that the interesting question is no longer “can it work” but “which of the nine ways fits what I am actually doing”.

A one-off photo of a receipt, a folder of 4,000 scanned invoices, and a screenshot you want to quote in a document are three different problems with three different right answers. This guide covers all nine, ranked by how quickly they get you to text.

Pick Your Method First

#	Method	Cost	Best for	Batch?
1	Phone camera	Free	One photo, right now	No
2	Google Drive + Docs	Free	A few files, editable output	Limited
3	Windows / macOS built-ins	Free	Screenshots on your desktop	No
4	Microsoft Office	Included	Documents you are already editing	No
5	Tesseract CLI	Free	Scripted local batches	Yes
6	Cloud OCR APIs	Per page	Large volumes, structured data	Yes
7	Adobe Acrobat	Subscription	Scanned PDFs, searchable output	Yes
8	Online converters	Free	Nothing sensitive	Limited
9	Vision language models	Per call	Messy layouts, handwriting, context	Yes

1. Your Phone Camera

The fastest path, and most people do not know it is built in.

iPhone (iOS 15+): Open the Camera or a photo, press and hold on the text. Yellow brackets appear around the recognised block. Drag to select, then Copy. It also works on text in the live camera view before you take a picture.

Android: Open Google Lens, either from the Assistant, the Camera app, or Google Photos. Point at the text, tap Text, then Copy text. “Copy to computer” pushes it straight to a signed-in Chrome desktop.

When This Is the Right Answer

• A business card, receipt, serial number, or WiFi password
• A page of a book you want to quote
• Anything where opening a laptop is more effort than the task deserves

Accuracy on clean printed text is excellent. It struggles with handwriting, dense multi-column layouts, and anything photographed at a steep angle.

2. Google Drive and Google Docs

The best free option for turning an image into an editable document rather than a clipboard string.

1. Upload the image or PDF to Google Drive
2. Right-click it → Open with → Google Docs
3. Drive runs OCR and produces a Doc with the original image at the top and the extracted text below it

What Makes It Good

• Handles multi-page PDFs
• Preserves basic structure: paragraphs, and often headings
• Free with any Google account
• Output is immediately editable and exportable to .docx

Limits

• Files up to 2 MB give the best results; large scans get downsampled
• Complex tables come out as loose text, not table structure
• Formatting fidelity is approximate, not exact

This is the method behind most “jpg to word” searches, and it is free.

3. Windows and macOS Built-Ins

Windows PowerToys → Text Extractor: Install Microsoft PowerToys, then press Win + Shift + T, drag a rectangle over any text on screen, and it lands on your clipboard. It works on text inside videos, error dialogs, and applications that block selection.

Windows Snipping Tool: Recent versions include Text actions after a capture, which extracts text and can redact email addresses and phone numbers.

macOS Live Text: Open any image in Preview, Quick Look, or Photos, and select text directly with the cursor. Works in Safari on images in web pages too.

When This Is the Right Answer

Anything already on your screen. There is no upload, no file to save, and no round trip.

4. Microsoft Office

OneNote: Paste an image into a page, right-click it, choose Copy Text from Picture. OneNote also indexes text inside images, so it becomes searchable in your notes.

Word: Insert a scanned PDF and Word offers to convert it to an editable document, running OCR as part of the conversion.

Excel: The Data → From Picture feature reads a photograph of a table and produces actual cells. It is the only method on this list that genuinely targets tabular data, and it asks you to confirm low-confidence cells before inserting.

If you are already inside Office, this is less friction than any external tool.

5. Tesseract

The open-source OCR engine, and the right answer when you have a folder rather than a file.

# Install
sudo apt install tesseract-ocr          # Debian/Ubuntu
brew install tesseract                  # macOS

# Single file
tesseract receipt.png output            # writes output.txt

# Straight to stdout
tesseract receipt.png -

# A whole folder
for f in scans/*.png; do
  tesseract "$f" "text/$(basename "${f%.*}")" 2>/dev/null
done

# Searchable PDF: the original image with an invisible text layer over it
tesseract scan.png output pdf

# Other languages (install the data pack first)
tesseract facture.png output -l fra
tesseract mixed.png output -l eng+deu

Page Segmentation Is the Setting That Matters

Tesseract’s accuracy depends heavily on telling it what shape the page is.

--psm	Meaning	Use for
3	Fully automatic (default)	Ordinary documents
4	Single column of variable sizes	Receipts, invoices
6	A single uniform block of text	Screenshots, cropped regions
7	A single line	Labels, serial numbers
8	A single word	Captchas, plate numbers
11	Sparse text, any order	Diagrams, signage

tesseract receipt.png output --psm 4
tesseract label.png - --psm 7

If accuracy is poor, change --psm before you change anything else. It is the single highest-impact setting.

Get Structured Output

# Tab-separated with per-word confidence and bounding boxes
tesseract scan.png output tsv

# hOCR: HTML with position data, useful for building overlays
tesseract scan.png output hocr

6. Cloud OCR APIs

When you have thousands of pages, or you need more than a flat string of text.

Service	Strength
Google Cloud Vision	Best general accuracy, strong on dense text and many languages
AWS Textract	Understands forms and tables as structures, not just text
Azure AI Vision	Good handwriting recognition, solid multi-language

# Google Cloud Vision
from google.cloud import vision

client = vision.ImageAnnotatorClient()
with open("invoice.png", "rb") as f:
    image = vision.Image(content=f.read())

response = client.document_text_detection(image=image)
print(response.full_text_annotation.text)

for page in response.full_text_annotation.pages:
    for block in page.blocks:
        print(f"block confidence: {block.confidence:.2f}")

Note document_text_detection rather than text_detection. The first is tuned for dense documents, the second for sparse text in photographs such as street signs. Choosing the wrong one costs real accuracy.

The Reason to Pay

Textract returns a form as key-value pairs and a table as rows and columns. Tesseract returns the same page as a stream of words you then have to reassemble yourself. If your output feeds a database rather than a human, that structure is the entire value.

7. Adobe Acrobat

The standard answer for scanned PDFs in an organisation that already has Acrobat.

Tools → Scan & OCR → Recognise Text. Acrobat keeps the original scan as the visible page and adds an invisible text layer behind it, so the document looks unchanged but becomes searchable, selectable, and accessible.

That output format matters. A searchable PDF is the correct archival form for a scanned document: you keep the visual record and gain the text, rather than choosing between them.

The free alternative produces the same structure:

# OCRmyPDF: adds a text layer to an existing PDF, keeps the original image
ocrmypdf --deskew --clean input.pdf output.pdf

# Force re-OCR on a file that already has a bad text layer
ocrmypdf --redo-ocr input.pdf output.pdf

8. Online Converters

Dozens of sites offer free image-to-text. They are convenient, and they carry a specific cost worth understanding.

The Privacy Question

Most of them upload your file to a server. For a photo of a poster, that is fine. For an invoice, a contract, a payslip, a passport, or a medical letter, you have just handed a document to a third party whose retention policy you have not read.

How to Choose One

Check	Why
Does processing happen in the browser?	If yes, nothing is uploaded
Is there a stated retention period?	“Deleted after 1 hour” is a real commitment
Is it HTTPS throughout?	Table stakes
Does it require an account for basic use?	Usually a data-collection signal

Our own browser tools run entirely on your device using WebAssembly, so files never reach a server. If the document is sensitive, prefer any local method: your phone, your operating system, Tesseract, or Acrobat.

9. Vision Language Models

The newest option, and the one that behaves differently from everything above.

Traditional OCR reads shapes and returns characters. A vision language model reads the image and understands it, so you can ask for the output you actually want:

"Extract every line item from this receipt as JSON with fields:
description, quantity, unit_price, total. Return only valid JSON."

Where It Beats Classic OCR

Situation	Classic OCR	Vision model
Clean printed page	Excellent, cheaper, faster	Overkill
Handwriting	Poor to fair	Much better
Messy multi-column layout	Confused reading order	Handles it
Table into structured data	Needs post-processing	Direct
“Which of these is the invoice total?”	Cannot answer	Can answer
50,000 pages	Cheap at scale	Expensive at scale

The Caveat That Matters

A vision model can hallucinate. Classic OCR fails visibly, producing garbage characters you can spot. A language model fails invisibly, producing a plausible number that was never on the page. For anything financial, legal, or medical, verify against the source or use classic OCR, where an error looks like an error.

Raising Accuracy on a Bad Source

Most OCR complaints are input problems, not engine problems. Five fixes, in order of impact.

1. Resolution

This is the one place in all of web imaging where DPI genuinely matters. OCR engines want roughly 300 DPI relative to the physical page, which for an A4 document means about 2,480×3,508 pixels. Below about 150 DPI, accuracy falls off a cliff.

If your scan is too low, upscaling before OCR sometimes helps:

magick scan.png -resize 200% -filter Lanczos upscaled.png

Everywhere else on the web, DPI is meaningless metadata. See the myths listicle for why.

2. Straighten It

Skew destroys line detection. A page rotated by three degrees can drop accuracy dramatically.

magick scan.png -deskew 40% deskewed.png
ocrmypdf --deskew input.pdf output.pdf

3. Increase Contrast, Then Binarize

OCR wants black text on white. Grey text on cream is much harder.

# Normalise, then threshold to pure black and white
magick scan.png -colorspace Gray -normalize -threshold 60% clean.png

# Gentler: adaptive threshold handles uneven lighting from phone photos
magick photo.jpg -colorspace Gray -lat 25x25-12% clean.png

The adaptive version (-lat) is the right tool for a photograph of a page, where one side is brighter than the other.

4. Remove Noise and Background

magick scan.png -despeckle -morphology Open Diamond clean.png
ocrmypdf --clean input.pdf output.pdf

5. Crop to the Text

Cropping away margins, staples, and the edge of the desk removes things the engine can mistake for characters.

magick photo.jpg -crop 1600x2200+120+180 +repage cropped.png

A Complete Preprocessing Pipeline

#!/usr/bin/env bash
# preprocess-for-ocr.sh — run before Tesseract on phone photos of documents
for f in "$@"; do
  out="prepped/$(basename "${f%.*}").png"
  magick "$f" \
    -colorspace Gray \
    -deskew 40% \
    -normalize \
    -lat 25x25-12% \
    -despeckle \
    +repage "$out"
  tesseract "$out" "text/$(basename "${f%.*}")" --psm 4
done

What Still Does Not Work Well

Be realistic about the limits so you do not waste an afternoon.

Input	Outlook
Cursive handwriting	Poor with classic OCR, fair with vision models
Heavily stylised display fonts	Poor
Text over a busy photograph	Poor
Very low resolution (under 100 DPI)	Poor, and upscaling rarely rescues it
Faded thermal receipts	Poor
Complex nested tables	Structure is lost without a form-aware service
Text at a steep perspective angle	Correct the perspective first

Perspective correction is worth doing before you give up:

magick photo.jpg -distort Perspective \
  '120,180 0,0  1480,140 1600,0  1520,2100 1600,2200  100,2140 0,2200' \
  flattened.png

OCR Is Not an Accessibility Fix

One important caveat. Running OCR over an image of text does not make that image accessible.

A screen reader does not run OCR. It reads the alt attribute. If your page contains a promotional banner with the offer baked into the pixels, the fix is to rebuild it with real HTML text, not to OCR it into an alt string.

<!-- Still inaccessible: zooming blurs it, translation cannot touch it -->
<img src="offer-banner.png" alt="25% off all lighting until 31 August">

<!-- Correct: real text over a decorative background -->
<div class="promo">
  <img src="promo-background.jpg" alt="" width="1200" height="400">
  <div class="promo__content">
    <h2>Summer sale</h2>
    <p>25% off all lighting until 31 August</p>
  </div>
</div>

The accessibility mistakes listicle covers this as mistake number one, and the screen reader guide covers the wider picture.

Summary

Choosing in One Line

You have	Use
One photo, on your phone	Live Text or Google Lens
Something on your screen	PowerToys Text Extractor or macOS Live Text
A few images, want an editable doc	Google Drive → Google Docs
A table in a photograph	Excel → Data → From Picture
A folder of scans, scriptable	Tesseract with the right --psm
Thousands of pages with forms	AWS Textract or Google Cloud Vision
A scanned PDF to archive	Acrobat or OCRmyPDF, keeping the image layer
Handwriting or messy layout	A vision language model, then verify
Anything confidential	A local method, never an upload site

Checklist

1. ✅ The source is 300 DPI relative to the page, or as close as you can get
2. ✅ The image is deskewed before OCR runs
3. ✅ Contrast is normalised, and phone photos use adaptive thresholding
4. ✅ --psm matches the page shape, if you are using Tesseract
5. ✅ Sensitive documents never go to an online converter
6. ✅ Scanned PDFs keep the image layer and gain a text layer, not one or the other
7. ✅ Vision model output is verified when the numbers matter
8. ✅ Nobody is using OCR as a substitute for accessible HTML text

Try the free method that matches your situation before reaching for anything paid. For most one-off jobs, the tool is already on the device in your hand.